I would recommend tightening your glasses
It's FP32 theoretical peak throughput, you can't cut it artificially in any way to differentiate products.
This whole thread is devoted to speculation. We can make reasonable inferences from AMD's slides and public statements, and from the gains that Nvidia obtained by implementing similar technology.
AMD told Anandtech that the MI25 Vega accelerator is going to have ~12.5 TFlops of single-precision performance. We know, on every GPU ever released in the modern era from both AMD and Nvidia, the formula for calculating TFlops. It's a simple function of (shaders x MHz x 2) / 1,000,000. Thus, if the flagship Vega chip has 4096 shaders (and all the information we've seen from multiple leaks so far indicate it does), then it has to run at ~1525 MHz to hit the TFlop goal. We also know based on past experience that professional cards almost never run higher clocks than consumer-grade cards; if anything, they run lower. From all this we can reasonably infer that the flagship consumer Vega card will run no slower than 1500 MHz, and quite possibly faster.
As for "IPC" - as used here, most people really mean something like "DX11 gaming performance per TFlop". Whether that is what AMD means in its architectural slides is not clear. And you're right that they didn't give specific numbers. However, again, we can make inferences from what we do know. AMD is implementing tiled rendering, so how much will that gain them? The closest comparison I can find is Nvidia's GTX 650 Ti (Kepler) to GTX 750 Ti (1st-generation Maxwell). Delta color compression (which both AMD and Nvidia now have), was, as far as I know, introduced only with 2nd-generation Maxwell (900 series), so using 1st-generation Maxwell as a baseline should mostly isolate the gains from tiled rendering. On 1080p gaming, GTX 750 Ti did about 22% better than GTX 650 Ti. This was despite the fact that the Maxwell card had only about 91.5% the raw TFlops of the Kepler card. In other words, Maxwell 1st generation had about 33% better DX11 gaming performance per TFlop than Kepler. (1.22 / 0.915) We can therefore reasonably infer that tiled rendering should provide significant gains for AMD in Vega, likely over 30% better DX11 gaming performance per TFlop. This is assuming no other significant bottlenecks - and we know from the slides that Vega does remove at least one major bottleneck, the 4-shader-engine limitation, which was one of the primary reasons that Fiji was so underwhelming.
These, then, are my predictions based on what we do know: clock speed for the flagship card of no lower than 1500 MHz (except perhaps in Furmark and other extreme TDP-constrained scenarios), and improvements in DX11 gaming performance per TFlop of no less than 30% over Polaris. We will see in a couple months whether I am right or wrong.
Do we know what process it will be on yet? Vega being on TMSC would be pretty promising newsThey're somewhat separate I think. Polaris was an Optimization of GCN 3 (or GCN 1.2 if you use that system), kind of like how Pascal isn't a truly new design either. Vega does have a lot more architectural changes than Polaris did. Polaris was the more conservative design to get the 14 nm process at GF worked out.
I don't know if the architecture itself clocks better, though I expect its something AMD will have worked on since the previous versions of GCN weren't the best at scaling to higher speeds. However, I think that one reason that Vega is so late is that AMD realized that the GF LPP process wouldn't let them hit the higher clock speeds (this should be blatantly obvious after looking at how quickly Ryzen OC falls off. Also Apple made their A9 using both TSMC's 16 nm and Samsung's 14 nm process, but the A9X which was clocked higher was only made at TSMC which adds further support to that notion) necessary so they shifted to using Samsung's newer LPU, which is supposed to be targeted for high performance. My guess is wwe get both an architecture and fab process that will enable Vega to hit higher clock speeds.
This assumes each CU is the same in terms of execution resources as prior iterations of GCN. Given how much they're talking about next gen CUs, I dont think we can assume the shaders are equal to prior versions of GCN. People keep making this implicit assumption. If that patent about having varying width execution units rolled into a single greater unit are true, then whatever is in the 64 CUs possibly (and likely, in my mind) wont put out the same number of FLOPS per clock. Thus dividing tflops by shaders to get frequency is flawed because we can't make that assumption vis a vis the shaders.
GCN is the name of of the uArch family. CU and NCU are the names of the shader array building blocks. NCU is not replacing the name GCN.Enjoying the discussion.
But I wanted to chime in. Remember, AMD uses different names than most of the tech media. Polaris is GCN 4. Vega is NCU. They consider this a big enough change to internally no longer refer to it as GCN.
But I view this with cautious enthusiasm. We should get higher clocks. We have the potential for higher performance-per-TFLOP. But with a new architecture I expect a lot of this potential left on the table with initial drivers and current games.
But anyway, it's fun to see the Vega-expectations-o-meter go from ~1080, to >1080 Ti, to somewhere in between based on who is posting. Once we get closer it'd be fun to have a Vega poll with where people think it will land. Not anonymous, of course.
Assuming that the GPU can't break ops apart or downgrade ops to a lower bit width opportunistically, or to be directed to do so in some capacity either by drivers or game code. If you could break some percentage of ops into smaller ones you could punch higher than your weight by increasing utilizationExactly the same FP32 op count per CU as with every GCN before it.
Exactly the same FP32 op count per CU as with every GCN before it.
GCN is the name of of the uArch family. CU and NCU are the names of the shader array building blocks. NCU is not replacing the name GCN.
Source?Vega is such big departure from GCN, that for compatibility Linux driver developers in 2015, were saying that the GPU would have to have abstraction layer, to create backwards compatibility with GCN GPUs. Vega requires completely fresh, and specific drivers for this architecture.
I think this will be apparent after the GPU launches.
I'm not saying GCN the name isn't dead, just that NCU isn't the replacement for it.You may be right, but do you have an official AMD source calling Vega GCN? They call Vega its own architecture in press releases. It's undoubtedly a bigger change than any previous iteration.
Maybe it's best to call it the Vega architecture, as AMD themselves do. We don't call Maxwell "Kepler 1.1". GCN as a name is dead.
An interesting post at SA forums:
While this does give some hope, it's also possible that utilisation isn't being reported properly, or that Vega has serious utilisation problems that aren't intentional sandbagging.
Good catch!
If true.... if it's doing 4k60 on a 3 out of 8 scale....
could be either good or bad, as you say, or simply AMD isn't using those LEDs as a power/usage meter anymore. I mean, there was the "Poor Volta" thing hidden in the "make some noise" Vega video...
We'll see tomorrow what AMD has prepared for Vega. I'll watch that live stream if I can.
Im sorry I will not find this post this year. So you can take this still with a grain of salt.Source?
I don't doubt that Vega is a big departure, but I haven't seen such comments from the Linux driver developers.
I'm not saying GCN the name isn't dead, just that NCU isn't the replacement for it.
So it IS a usage meter just like on Fury X....
Do we know what process it will be on yet? Vega being on TMSC would be pretty promising news
Source?
I don't doubt that Vega is a big departure, but I haven't seen such comments from the Linux driver developers.
I'm not saying GCN the name isn't dead, just that NCU isn't the replacement for it.
Doesn't the implication of the move to NCU suggest a fairly large departure. We're obviously not going to see a completely new radical design for a GPU, but it seems like between this and their HBCC they've done more than just the usual incremental improvement. Regardless of what they call it, it seems like a big difference compared to the previous generation.
I already know that, and made the same arguments in the past. I'm just curious about the linux driver team stating anything of the sort.I don't doubt that Vega is a big departure, but I haven't seen such comments from the Linux driver developers.
I already know that, and made the same arguments in the past. I'm just curious about the linux driver team stating anything of the sort.
He said none of these things and it is clear that you are blind to everyone's comments. You are a pure troll, and it astonishes me that you keep posting this nonsense.
You want to know what is astonishing? Your inability to read with understanding of what people are writing, and your projection of your perception on others words.What is astonishing, is you lying to everyone in this this whole thread, suggesting He didn't say those things (in which he clearly has), then calling me a troll for pointing his bias out.
What does that say about you..?
He has repeatedly suggested, there is not enough information out there, therefore AMD sucks. He has said, he can't even speculate (or imagine) that AMD tech could be better? And is backing up his remarks by further suggesting, someone to knock on his door & hand him proof (Because Nvidia Yo..)
Instead of connecting the dots and understanding the backstory to Vega.
What is astonishing, is you lying to everyone in this this whole thread, suggesting He didn't say those things (in which he clearly has), then calling me a troll for pointing his bias out.
What does that say about you..?
He has repeatedly suggested, there is not enough information out there, therefore AMD sucks. He has said, he can't even speculate (or imagine) that AMD tech could be better? And is backing up his remarks by further suggesting, someone to knock on his door & hand him proof (Because Nvidia Yo..)
Instead of connecting the dots and understanding the backstory to Vega.